[BUG] unexpected group order change in VariableImportance (#632)

* add regression test

* fix group order

* update changelog

* docsrtring

Author: jpaillard

CHANGELOG.rst

@@ -19,6 +19,7 @@ Bug fixes
 ---------
 
 * Fixed a bug in CLuDL/EncluDL: cloning the internal estimators to avoid them carrying unwanted history. (PR #425)
+* Fix the unexpected change in the order of feature groups when a dataframe is used as input (PR #632)
 
 Maintenance
 -----------

src/hidimstat/base_variable_importance.py

@@ -541,14 +541,14 @@ def fit(self, X, y=None):
                 j: [j] for j in range(self.n_features_groups_)
             }
             self._features_groups_ids = np.array(
-                sorted(self.features_groups_.values()), dtype=int
+                list(self.features_groups_.values()), dtype=int
             )
         elif isinstance(self.features_groups, dict):
             self.features_groups_ = self.features_groups
             self.n_features_groups_ = len(self.features_groups_)
             if isinstance(X, pd.DataFrame):
                 self._features_groups_ids = []
-                for features_group_key in sorted(self.features_groups_.keys()):
+                for features_group_key in self.features_groups_:
                     self._features_groups_ids.append(
                         [
                             i

test/test_base_variable_importance.py

@@ -1,10 +1,11 @@
 import matplotlib.pyplot as plt
 import numpy as np
+import pandas as pd
 import pytest
 from sklearn.cluster import FeatureAgglomeration
-from sklearn.linear_model import LassoCV
+from sklearn.linear_model import LassoCV, LinearRegression
 
-from hidimstat import CluDL, DesparsifiedLasso
+from hidimstat import CFI, CluDL, DesparsifiedLasso
 from hidimstat.base_variable_importance import BaseVariableImportance
 from hidimstat.statistical_tools.multiple_testing import fdp_power
 from hidimstat.statistical_tools.p_values import two_sided_pval_from_pval
@@ -527,3 +528,26 @@ def test_clustered_fwer_selection(rng):
         fwer=0.5, n_tests=n_features, two_tailed_test=True
     )
     assert selection.shape[0] == n_features
+
+
+@pytest.mark.parametrize(
+    "n_samples, n_features, support_size, rho, seed, value, signal_noise_ratio, rho_serial",
+    [(100, 10, 5, 0.5, 0, 1.0, 64.0, 0.5)],
+)
+def test_feature_groups_order_preserved(data_generator):
+    """Regression test to check that the order of the feature groups is
+    preserved in the output of .importance() even when the keys are not in
+    the order that would be returned by sorted(groups.keys()).
+    """
+    X, y, important_features, non_important_features = data_generator
+    X_df = pd.DataFrame(X)
+    model = LinearRegression()
+    model.fit(X_df, y)
+    groups = {
+        "non_important": non_important_features,
+        "important": important_features,
+    }
+
+    cfi = CFI(estimator=model, features_groups=groups, random_state=0)
+    importance = cfi.fit_importance(X_df, y)
+    assert importance[0] < importance[1]